Deep Neural Network Based Disease Discrimination Learning from Small Medical Image Training Set and User Feedback

Deep learning has a good prospect in medical diagnosis, and training a disease discrimination model requires a large number of samples. However, in many cases, it is difficult to achieve a good performance due to the limitation of the real training samples. This paper proposes a novel small sample expansion (SSE) method to generate new images for deep learning from a small medical image training set. In SSE, we employ user feedback to mark very limited amount of training images, and highlight and transform the feature of key areas to expand the training set automatically. We take the diagnosis of trichocryptosis as example, experimental results show the improvement of accuracy rate and conviction rate of the proposed method.